Every turn of an AI conversation resends and recharges the whole history. On long sessions this context re-read is often the largest line on the bill.
AI models do not remember a conversation between requests. Each time you send a new message, the application sends the entire conversation again: every earlier question, every earlier answer, every file that was attached. The provider charges for all of it as input. A session that started cheap becomes expensive turn by turn, simply because the history keeps growing.
For agents and coding assistants, which can run dozens of steps over a large codebase, this context re-read is often the biggest line on the AI bill.
Providers know this, and most offer a large discount on history they have recently seen. When the same model receives the same conversation prefix again, the repeated part is charged at a fraction of the normal rate. Keeping a long conversation on the model already holding it can cut the re-read charge to about a tenth.
That discount is lost the moment the conversation moves to another model. The new model has never seen the history, so it is charged in full.
The right choice depends on the numbers. Sometimes a different model is cheaper enough per token that switching pays for itself even after losing the discount. Often it is not. The only way to know is to price both paths before the next request is sent.
CompletionPrism™ counts context re-read and switching as two of the five costs it prices before every step. When moving the conversation would cost more than it saves, it stays on the current model and the receipt says why. When a move is worth it, it makes the move and shows the trade.
Look at your longest AI sessions from last week. Compare input spend with output spend. If input is many times larger, most of what you paid for was the same history sent again.